Abstract

Background/Aim: Conjunctivitis, commonly known as pink eye, poses a significant diagnostic challenge due to overlapping presentations across viral, bacterial, and allergic subtypes. This study presents MOGONET, a multi-objective generalized normal distribution optimization framework built upon VGGNet, specifically designed to address conjunctivitis identification through resource-aware deep learning optimization.
Methods: The proposed pipeline integrates contrast-limited adaptive histogram equalization-based image enhancement, data augmentation, and multi-level Otsu thresholding segmentation with transfer learning using a VGG16 backbone. A MOGNDO-based outer-loop search simultaneously maximizes classification accuracy, minimizes binary cross-entropy loss, and minimizes computational cost (FLOPs) using Pareto-front-based non-dominated sorting. Experiments were conducted on 1,230 conjunctival eye images sourced from Kaggle and Shutterstock, with a stratified 72%/18%/10% train/validation/test split.
Results: MOGONET achieves 98.32% test accuracy, 97.54% precision, 96.92% recall, and 96.22% F1-score. Compared with the unoptimized Visual Geometry Group-based baseline, MOGONET reduces FLOPs by 79.18% (from 238.06M to 49.56M), and model parameters by 77.78% (from approximately 9M to 2M).
Conclusion: MOGONET demonstrates that multi-objective optimization can substantially reduce computational complexity in conjunctivitis classification without sacrificing diagnostic performance, offering a promising framework for deployment on resource-constrained clinical screening platforms.

Key words: Conjunctivitis, Generalized Normal Distribution Optimization, multi-thresholding, deep learning